Granite-3.1-2B-Instruct (dense) with layers 19-23 duplicated. A mid-stack reasoning circuit runs twice on every forward pass.

40 base layers → 45 after duplication. No training, no merging, no weight changes.

The dense-Granite control. This is the within-family counterpart to the MoE sibling Granite-3.1-1B-A400M-RYS-12-15. The contrast is sharp: the MoE lifted reasoning +52.94% but degraded EQ on every config (−13.52 at the best-combined pick). The dense 2B lifts reasoning a modest +11.76% but holds EQ stable (+2.50) — with the trade-off appearing instead between math and reasoning (math −9.86). MoE makes RYS more aggressive in both directions; dense moves a smaller trade across a different axis.

The largest swept configuration count in the small-model queue (87 configs) makes this the most thoroughly characterized dense response in the corpus. Pick this when you want a reasoning lift with stable EQ and can absorb math cost.

87 configurations tested. (19,24) block-5 is the best-combined pick. Full per-config sweep + cross-architecture analysis: v2 dataset.

Part of the RYS Sovereign Collection v2.

v1 — Qwen2.5 cross-scale + Qwen3-32B headline crossover. 5 model repos.

v2 — cross-architecture corpus. 21 model variants across 10 architecture families. Inverse correlation (r = −0.726): weak baselines lift more, in their weakest dimension. The Granite-3.1 family (MoE 1B-A400M + dense 2B) is the cleanest MoE-vs-dense mechanism contrast in the corpus. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.

Within-family sibling: john-broadway/Granite-3.1-1B-A400M-RYS-12-15-GGUF — the MoE that lifts reasoning hard but pays in EQ.

John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 sweep generation and build pipeline; Opus 4.7 in May 2026 cross-architecture analysis and publication). Original RYS method by David Ng on Qwen2-72B; sweep + probe toolkit by alainnothere.